Anthropic's S-1: $80.7B half-year loss and up to $750B in compute commitments as Oracle CDS hits a record 343bp

The outline of Anthropic's record IPO is emerging — and credit markets have priced it in advance. Per the Wall Street Journal, citing the latest draft S-1, the company lost more than $80.7 billion net in the first half of 2026 and has locked in compute commitments of up to $750 billion; the same day, Oracle CDS — the weather vane of AI debt risk — hit a record 343 basis points intraday, with Meta's and SpaceX's CDS also at records. However heavy the AI bill, the market is voting with the price of default insurance^[1][2].

Three sets of numbers in the S-1

The first is the loss. H1 2026 net loss exceeded $80.7 billion, of which roughly $72 billion was non-cash accounting accrual from revaluation of convertible notes and similar instruments — "paper loss"; stripping that out, operating losses ran about $8.7 billion, already above the full-year 2025 total of over $8 billion. The core driver: near-$9 billion of direct compute and infrastructure spend in H1, almost four times the $2.3 billion a year earlier — before paying any salaries, more than 55 cents of every revenue dollar already flows to datacenters and chip suppliers^[1]. The second is the revenue-valuation mismatch. H1 revenue was $16.3 billion, with Q3 up 45% sequentially to $16.8 billion — but sequential growth has decelerated sharply from ~140% in Q1-Q2 to 45–46%, prompting some institutions to cut reference valuations from $2 trillion to about $1.5 trillion^[1]. The third is the heaviest: compute commitments. Locked infrastructure commitments have breached $600 billion, reaching $750 billion if the additional $150 billion under negotiation (including Nscale) lands; of that, about $450 billion over the next seven to ten years is irrevocable cloud obligations — $111 billion to Google Cloud, $110 billion to AWS, $31 billion to Microsoft, up to $86 billion in chip leases to Nscale and Lambda running to 2033–2034 — plus over $160 billion of TPU lease obligations via Broadcom, some of it financed off-balance-sheet through SPVs and private credit^[1][3]. The S-1 also discloses a supply-side bind: Anthropic's demand already exceeds what the three hyperscalers will sell it — they increasingly see it as a direct competitor — forcing it to build its own datacenters and partner with operators of "limited operating history"^[1].

Credit markets: three CDS arrows at once

The same week the S-1 details surfaced, credit markets reacted violently. On Tuesday Oracle CDS hit a record 343bp intraday; Meta's CDS gauge rose 6.3bp to a record 114.8bp; SpaceX five-year CDS touched an intraday record 242bp — more than double the ~111bp when trading began in June^[2]. The driver is supply-shock expectations: with earnings season ending, blackout periods lift and the market expects a fresh wave of mega bond issuance — Meta, which sold $25 billion in six tranches in April after hinting it wouldn't return before Q4, is now widely expected to tap the dollar market again to fund AI spending, and is talking to European investors about a potential euro deal; BofA data showed dealers net-buying about $451 million of hyperscaler bonds Monday (the most since mid-September), read as investors making room by selling^[2]. The scale of hedging demand is striking: per Barclays, CDS volume tied to the ten largest hyperscalers and chipmakers soared 691% in the first nine months to a record $140 billion (from about $17.7 billion a year earlier); global corporates have issued over $450 billion of AI-funding bonds this year, and SpaceX is negotiating a $40 billion financing^[2].

The bubble debate: Griffin's "Amazon" answer

Against this data, the bubble debate was bound to heat up. Citadel founder Ken Griffin told his firm's annual conference on Tuesday that AI is "almost certainly" in a bubble — but offered a historical frame: "any fundamental transformation of the economy is usually intertwined with a bubble," and the companies that survive the other side become market leaders for decades — "I just wish I hadn't sold my Amazon shares"^[4]. His arc is notable: "garbage" at Davos in January, "frustrated" at Stanford in May, now decidedly more optimistic and predicting a golden age for entrepreneurs; he also cautioned that many companies are re-examining what they pay Anthropic and OpenAI, and that Anthropic's success ultimately hinges on proving its models are "a necessity" in every industry^[4].

What to watch

First, Meta's post-earnings bond moves — a dollar-plus-euro double tap would formally open the Q4 AI issuance wave. Second, Anthropic's public filing in the coming weeks, where the $1.5T-vs-$2T valuation fight puts "paper loss versus real cash burn" pricing squarely on the table. Third, knock-on effects in high-beta credits like CoreWeave. Fourth, whether the Fed's rate path keeps pressuring long-duration AI bond yields^[1][2][4].